Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation

arXiv:2610.22715v1 Announce Type: cross Abstract: Refusal is a common safety mechanism in large language models (LLMs), yet in contexts of suicidal ideation (SI), refusing to engage can itself constitute a safety failure. We evaluate how eight instruction-tuned LLMs respond to SI and self-harm disclosures across three datasets, varying whether models are informed that a disclosure expresses SI and whether they receive a clinician-developed response protocol, and complement this with mechanistic analyses in six open-weight models. Our results reveal a nuanced relationship between refusal and response-level risk. At baseline, elevated risk in model responses is driven primarily by inattention to the disclosure, alongside refusal to engage. Mechanistic analyses suggest that models may refuse SI disclosures through the same direction that mediates their refusals of harmful requests. However, informing models of the SI context, or suppressing refusal directly, increases engagement without reliably lowering response risk, and may raise it. In contrast, the clinician-guided protocol consistently reduces response-level risk while also reducing refusal. Together, these analyses and ablations motivate the need for models that refuse only when warranted and engage in interactionally safe ways when users disclose SI.

LLMs may refuse suicidal ideation disclosures, increasing risk. Informed models or suppressed refusal may raise risk.

The study identifies that refusing to engage with suicidal ideation disclosures by LLMs can itself be harmful, suggesting current refusal mechanisms may fail in these contexts. However, clinician-guided protocols show promise in reducing risk and refusal, indicating potential for safer engagement. The implications are mixed due to risks of harmful responses and opportunities for improved interventions.

This is a rigorous academic study showing current LLM refusal behavior can worsen outcomes for suicidal users, a concerning finding about deployed systems. However, it also identifies a working mitigation (clinician-guided protocols), making the overall implication mixed rather than purely negative.

LLMs at baseline exhibit risky behavior when responding to suicidal ideation, either by refusing to engage or by providing unhelpful responses. Attempts to increase engagement without specific protocols increase risk. This highlights a significant safety failure with potentially severe consequences for human well-being.

The study shows that current LLM safety mechanisms can fail in contexts of suicidal ideation. While there are ways to improve model responses, the baseline risks are concerning.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   74a49f72-5a37-47c5-92d7-865597bee500
           https://arxiv.org/abs/2610.22715
   ↓
EVIDENCE   3 extracted excerpts
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -43  (Adverse)
   ↓
CONFIDENCE 82%